self-improve-bot

name	self-improve-bot
description	Train and verify Cyberpunk greedy-strategy weights with the ai-runner hill climber. Use when changing greedy automation, trainer behavior, bot heuristics, or evaluating a candidate weight set.

Self-Improve Cyberpunk Bot

Before changing game behavior, load cyberpunk-tcg-rules. Then inspect:

• packages/engine/src/automation/README.md
• packages/engine/src/automation/strategies/greedy.ts
• tools/ai-runner/src/train.ts
• tools/ai-runner/src/cli.ts

Contract

The trainer mutates GreedyWeights, evaluates each candidate against a fixed opponent and seed batch, and accepts only strictly better win rates. Training results are candidates, not proof: promote weights only after a fresh-seed comparison against the default strategy.

Workflow

1. Run focused automation and trainer tests. Record unrelated baseline failures; do not silently classify warnings as safe.

bun test packages/engine/tests/automation tools/ai-runner/tests

2. Choose an opponent with headroom. Avoid a saturated matchup where the baseline already wins nearly every game.
3. Train with a stable seed and explicit output path:

node --experimental-transform-types tools/ai-runner/src/cli.ts train \
  --opponent greedy \
  --matches 20 \
  --iterations 40 \
  --seed self-improve-<run-id> \
  --output tools/ai-runner/trained/weights-<run-id>.json

Use --real-cards only when the requested evaluation should use the real card catalog. Read tools/ai-runner/src/real-catalog.ts before claiming what deck distribution that flag represents.

4. Confirm the output schema contains bestWeights, baseline and best win rates, opponent, match count, iterations, and seed.
5. Compare the trained strategy with the default on a different seed batch. Use enough games to distinguish a stable gain from one or two game flips.
6. Treat a tie or fresh-seed regression as no demonstrated improvement. Increase the match count or broaden the evaluation; do not promote the weights.
7. Re-run focused engine tests after any default-weight or strategy change.

Promotion Gates

Ask before committing generated weight artifacts or changing user-facing bot strength. There are two separate integration surfaces:

• Engine default: packages/engine/src/automation/strategies/greedy.ts
• Practice UI adapter: ../agnostic-simulator/apps/multi-game-simulator/src/games/cyberpunk/engine/practice/practiceEngine.ts

The practice adapter currently resolves registered strategies through getSafeAutomatedActionStrategyOption. Re-read that code before deciding whether a trained strategy belongs in the engine registry or the adapter; do not replace it with an obsolete hard-coded greedyStrategy recipe.

Report

Include the training seed and parameters, baseline and best training results, fresh-seed trained-versus-default result, output path, focused checks, and a clear promote or do not promote conclusion.